Automatic Evolutionary Clustering for Human Activity Discovery

Clustering is regarded as a good approach to distinguish between different human activities from skeletal data in an unsupervised manner (also known as human activity discovery) because it does not require the laborious task of labeling a huge volume of data. In this chapter, we demonstrate a multi-objective evolutionary clustering methodology using particle swarm optimization, game theory, and Gaussian mutation techniques for performing such a task. The proposed methodology does not require any parameter setting nor prior knowledge of the number of clusters. It uses an automatic segmentation method based on kinetic energy to reduce redundant frame and identify keyframes. Features that characterize human motion are extracted from these keyframes and their dimensions are reduced using principal component analysis (PCA) before performing clustering on the reduced dataset. The proposed methodology was tested on popular benchmark datasets such as Cornell activity dataset (CAD-60), Kinect activity recognition dataset (KARD), Microsoft Research (MSR), Florence3D (F3D), and Nanyang Technological University (NTU-60) and compared with four automatic and four nonautomatic clustering algorithms, outperforming the other algorithms in most datasets. We demonstrate that the application of game theory enabled our clustering methodology to find the global best which is the optimal solution based on the multi-objective functions. We also showed that our methodology converges quickly due to the effects of game theory and Gaussian mutation.